Meta released Muse Wonder on 12 October 2026 as the API-accessible half of a 2-model line, with 6 reasoning levels from minimal to max. It suits developers who want Meta's agentic model for long coding and tool-use jobs without a Muse Code subscription, and who accept that Muse Wonder Full scores higher.
What changed
- · Release Meta releases Muse Wonder on the Meta Model API and in Muse Code (muse-wonder-1.0) source
Muse Wonder scores 52 on the Artificial Analysis Intelligence Index v4.3.2 and is Meta's proprietary reasoning model for long coding and agent work. It accepts text, image, video, audio and PDF input, returns text, has a million-token context window, and can be reached through Meta's own API or Muse Code.
Where it sits
- $8.00/M$ per 1M tokensBlended price (3:1)Lower is better#63 / 80peer median $1.86/Mvendor price, checked by HokAI
- 131 tok/stokens/sOutput speedHigher is better#18 / 50peer median 91 tok/scited: Artificial Analysis
Pricier than 80% of the 80 GA models with a published price, rank 18 of 50 on output speed as cited from Artificial Analysis, and one of 77 whose vendor states it does not train on customer data. Ranked against GA models; this record is not GA.
Ranks are against GA models on HokAI that publish the same figure; ties share a rank.
Provider: Meta AI · Family: Muse
Context window: 1,048,576 tokens
Input modalities: text, image, video, audio, pdf, tool-calls · Output: text, tool-calls
About Muse Wonder
Muse Wonder is a proprietary reasoning model from Meta AI, released on 12 October 2026 together with Muse Wonder Full. It sits above Muse Spark 1.3 in the Muse family (see all Meta models), and Meta's docs steer long, complex jobs such as multi-file coding changes, multi-hour agent runs and sub-agent orchestration to it, leaving high-volume or latency-sensitive work to Spark. Meta trained the pair at its gigawatt-scale Prometheus data center and has not published a parameter count, architecture type or training cutoff. The Meta Model API docs still tag Muse Wonder as a preview.
On benchmarks the picture is strong for coding and mixed elsewhere. Meta's comparison table, run at max reasoning effort, lists Muse Wonder at 78.8 on DeepSWE v1.1, a long-horizon coding test where the figure comes from a third-party Datacurve run. Meta lists Claude Opus 5.5 at 74.2 and GPT-6 Astra at 74.1 on the same row, which puts Wonder ahead of both. It trails both on Terminal-Bench 4.0, scoring 52.0 against 59.6 and 59.1, and Artificial Analysis independently puts it at 52 on its Intelligence Index v4.3.2, below the 58 that Meta's table lists for Opus 5.5. Meta says it takes the highest of its own validated runs, self-reported scores or leaderboard results for each row, so read the table as vendor-reported and check Artificial Analysis for the independent view.
Inputs cover text, images, video, audio recordings and PDF files, with text as the only output. Context runs to 1,048,576 tokens, and the docs list tool calling, structured output, search grounding and computer use that can drive browsers, desktop and mobile apps. Reasoning runs at one of six levels from minimal to max, and switching reasoning off entirely is not supported. Meta's model card names visual and spatial reasoning, especially 3D tasks, as a known weak spot, and says the model was not evaluated on every language in its pretraining data.
You can call it on the Meta Model API with an OpenAI-compatible client through the Responses, Chat Completions or Messages protocols, inside Muse Code, Meta's terminal coding agent and a rival to Claude Code and OpenAI Codex, or through OpenRouter, where Meta is the only listed provider. Meta's docs list no Amazon Bedrock, Google Vertex or Azure route. Rate limits apply per team rather than per key. Meta also says Muse Wonder will soon power experiences in the Muse agent app, rolling out first to subscribers.
Meta says it trained the Wonder family to avoid sharing private information without permission and to turn down clearly harmful requests, under its Meta Superintelligence Scaling Framework. The model card's risk section, written about Muse Wonder Full and applied to the whole family, says Meta cannot rule out the high-risk threshold in chemical and biological misuse without mitigations, so the models ship with safety post-training and system-level classifiers. With those in place Meta rates residual risk as moderate or lower in every domain it assessed. In the API's Standard mode your prompts, code and responses are not used for training.
Choose it when capability matters more than cost per token and you want to pay by usage rather than by subscription: large refactors, long agent loops and computer-use workflows. Skip it for latency-sensitive or high-volume traffic, where Spark is cheaper, and look at the subscription option if you want the best Meta result, which Meta keeps behind a Muse Code plan. Buyers comparing across vendors should also weigh Claude Fable 5.1 and, when cost per token dominates, DeepSeek-V4-Pro or Qwen3.8-Max. Then read the guide to choosing an LLM and the AI coding assistants roundup.
Pricing
Meta Model API, Standard mode: $4.00 per 1M input tokens, $20.00 per 1M output tokens and $0.20 per 1M cached input tokens, with the same rates whether the context is nearly empty or nearly full. Reasoning tokens bill as output. The web search tool is charged separately at $2.50 per 1,000 queries. No batch rate is listed, and the discounted Contributor mode is unavailable for this model on the API. Muse Wonder is also included in every Muse Code plan; the flat-rate top tier that adds [Muse Wonder Full](/hub/models/muse-wonder-full) is priced on that page. Compare against [Muse Spark 1.3](/hub/models/muse-spark) when cost per token is the constraint.
What a real job costs
| Job | Input | Output | Total |
|---|---|---|---|
| Summarise a 20-page PDF | $0.120 | $0.020 | $0.140 |
| Support reply | $0.0080 | $0.0060 | $0.014 |
| One coding agent run | $0.800 | $0.400 | $1.20 |
Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.
Key Features
- Six reasoning levels: Set reasoning_effort from minimal to max on the Responses or Chat Completions protocol; none is rejected with HTTP 400.
- 1,048,576-token context window: Takes roughly a million tokens per request, billed at the same per-token rate whether the window is nearly empty or nearly full.
- Five input types: Accepts text, image, video, audio and PDF in the same request and returns text.
- Parallel tools and sub-agents: Meta's docs describe planning long tasks, delegating parts to sub-agents, calling tools in parallel and recovering from errors across many steps.
- Computer use: Agents can see and operate browsers, desktop and mobile apps; Meta ships a cookbook that gives the model screen access.
- Caching and search grounding: Cached prefixes bill at a lower input rate, and the web_search tool adds live results for a separate per-query fee.
- Team-level rate limits: 3,000 requests and 4,000,000 tokens per minute per team, counted across every API key.
Pros
- Within 0.2 points of Muse Wonder Full on DeepSWE in the Datacurve run, and the only Wonder model you can buy by the token.
- Lists at the same per-token price as Claude Opus 5.5 on HokAI's records, and at under half the price of Claude Fable 5.1 or GPT-6 Astra.
- Meta's internal MemoryBench puts it marginally ahead of Full and ahead of every rival in Meta's table, a vendor-run test worth confirming on your own data.
- Meta's docs say moving a workload between Muse Wonder and Muse Spark is a model ID change on the same endpoints.
Cons
- Scores below Muse Wonder Full on every row of Meta's table except MemoryBench, and Full cannot be bought by the token.
- Artificial Analysis logged 130M output tokens running its index, so cost per task depends on the effort setting more than the per-token rate suggests.
- Still labelled a preview in Meta's docs, with no Bedrock, Vertex or Azure route listed.
- Trails Claude Opus 5.5 on terminal tasks and on building programs from scratch in Meta's own table.
Benchmarks
- AA Intelligence Index: 52 cited: Artificial Analysis · 14 Oct 2026 — Composite of 10 evaluations run by Artificial Analysis, 0 to 100.
- Output speed: 131 tok/s cited: Artificial Analysis · 14 Oct 2026 — Median tokens written per second as measured by Artificial Analysis.
A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.
Frequently Asked Questions
How much does Muse Wonder cost in 2026?
The Meta Model API charges $4.00 for every million tokens in, $20.00 for every million tokens out and $0.20 for every million cached tokens. Meta states there is no long-context premium, reasoning tokens bill as output, and web search grounding adds $2.50 per 1,000 queries. Meta lists no batch rate and no discounted Contributor mode for this model. On HokAI's records Claude Opus 5.5 carries the same per-token prices, while Claude Fable 5.1 and GPT-6 Astra list at $10.00 and $50.00.
Is Muse Wonder better than Claude Opus 5.5?
It depends on the job. In Meta's own table Muse Wonder leads on DeepSWE and SWE-Marathon, but Opus 5.5 leads on Terminal-Bench 4.0, ProgramBench and the Artificial Analysis index, so Wonder is the pick for long repository work and Opus 5.5 is the safer pick for terminal tasks and from-scratch builds. The SWE-Marathon row mixes different coding tools, and Meta chose which results to show, so check Artificial Analysis before relying on the gap.
Is Muse Wonder open source or proprietary?
Proprietary. Meta serves it only through its own API, Muse Code and OpenRouter, and Artificial Analysis lists it as a proprietary model. The open-weight model in Meta's lineup is [Muse Glimmer](/hub/models/muse-glimmer), which runs on your own hardware, so Wonder cannot be self-hosted. Use is governed by the Meta Model API Terms of Service and Acceptable Use Policy.
Does Meta train on what you send to Muse Wonder?
Not in Standard mode, the only mode the Meta Model API offers for this model: Meta's docs say prompts, code, files, context and responses are not used to train Meta models. Muse Code subscribers can use Contributor mode, which is on by default on subscriptions and lets Meta train on that content in exchange for more usage; it can be switched off in settings. The pages Meta publishes for the model do not state a retention period or a zero-data-retention option.
Who should use Muse Wonder, and who should skip it?
Use it for multi-file coding changes, multi-hour agent runs and computer-use workflows where you want to pay by usage. Skip it for high-volume or latency-sensitive traffic, where Muse Spark 1.3 is the cheaper route Meta itself recommends, and skip it if you need Bedrock, Vertex or Azure deployment, which Meta's docs do not list. Teams that want Meta's strongest result need a Muse Code subscription for Muse Wonder Full.
Top Alternatives
- Claude Opus 5.5: Pick Muse Wonder for long repository coding runs; pick Claude Opus 5.5 for terminal tasks and from-scratch programs, where Meta's own table has it ahead.
- GPT-6 Astra: Pick Muse Wonder when per-token cost matters; pick GPT-6 Astra if the higher terminal and ProgramBench rows in Meta's table decide it.
- Claude Fable 5.1: Pick Muse Wonder for coding depth at a lower token price; pick Claude Fable 5.1 if the computer-use rows in Meta's table matter more.
- Muse Spark 1.3: Pick Muse Wonder for hard, long jobs; pick Muse Spark 1.3 for high-volume or cost-sensitive work, as Meta's docs advise.
- Muse Wonder Full: Pick Muse Wonder for pay-per-token API access; pick Muse Wonder Full if you hold a Power or Maximum Muse Code plan and want Meta's top scores.